Multi-agent autonomous cooperative control method at the edge of power distribution network

Abstract

This invention discloses a multi-agent autonomous collaborative control method for distribution network edges. To address the problems of high response latency and poor interoperability of heterogeneous devices in traditional centralized architectures, this invention deploys a lightweight large language model on edge device nodes, constructing a semantic primitive communication layer containing "intent-parameter-constraint" triples. When a trigger event occurs, the source device agent broadcasts a semantic primitive message; the receiving agent parses the message and calculates the adjustment bid using its local model under secure hard constraints; the source device agent generates candidate adjustment schemes through multi-round bidding negotiation; and the schemes are executed after passing mandatory verification by an independent security verification module. This invention reduces response latency to milliseconds, achieves semantic interoperability and plug-and-play functionality for heterogeneous devices, and effectively manages the risk of decision illusions in large language models through dual protection of secure embedding at the model layer and independent verification at the decision layer. It is suitable for real-time control scenarios at the edge of distribution networks with a high proportion of distributed energy access.

Description

Technical Field

[0001] This invention relates to the field of edge computing and intelligent control technology for power distribution networks, and in particular to a multi-agent autonomous collaborative control method for power distribution network edges. Background Technology

[0002] With the deepening implementation of the "dual-carbon" strategy, my country's power distribution network is undergoing a fundamental transformation from a unidirectional, passive network to a multi-source, collaborative active network. The scale of distributed photovoltaic, electrochemical energy storage, electric vehicle charging piles, and various flexible loads is growing exponentially. The new pattern of highly integrated "source-grid-load-storage" has completely broken the traditional power distribution network's operating mode of "unified dispatching on the generation side and passive consumption on the consumption side," posing unprecedented challenges to the communication and control system of the power distribution network.

[0003] Traditional power distribution networks adopt a centralized communication and control architecture of "master station-terminal". The operating data of all edge devices must be uploaded to the power distribution network master station in real time. The master station performs global unified calculation and decision-making before issuing control commands to each terminal. In scenarios involving large-scale distributed energy integration, this architecture exposes inherent and irreconcilable flaws: First, the response latency cannot meet real-time control requirements. The end-to-end latency of traditional centralized control is typically between 100ms and 1000ms (milliseconds to seconds), while the time constant of photovoltaic power surges is only 10ms to 50ms, and the protection response requirement for voltage dips is strictly less than 20ms. The response speed of the centralized architecture is at least an order of magnitude slower than the emergency control requirements, which can easily lead to voltage overruns, frequency fluctuations, or even large-scale power outages. Second, the communication bandwidth and security pressure increase dramatically. Uploading massive amounts of data to the main station not only occupies more than 90% of the communication bandwidth but also increases the security risk of data theft and tampering during transmission. Third, the interoperability of heterogeneous devices is extremely poor. There are dozens of mainstream communication protocols in the power industry, and different manufacturers and models of equipment use different protocol standards. The protocol configuration cycle for a complex distribution area can take as long as 4-6 weeks, resulting in extremely high operation and maintenance costs.

[0004] To address these issues, academia and industry have proposed a multi-agent cooperative control approach, deploying agents on edge devices to offload some decision-making capabilities to the edge. However, existing multi-agent cooperative solutions still suffer from fundamental flaws: the communication layer remains constrained by traditional protocols, and interactions between agents still rely on traditional power protocols based on binary messages. Agents developed by different manufacturers cannot communicate and collaborate directly, rendering the multi-agent system a "pseudo-distributed" system, essentially still multiple independent centralized subsystems. Furthermore, existing distributed negotiation algorithms suffer from high computational complexity and slow convergence, typically requiring several seconds or even tens of seconds to reach consensus, failing to meet the millisecond-level real-time control requirements of power distribution networks. More critically, most multi-agent systems still retain the role of the master station or edge gateway as the central coordinator; if the central node fails, the entire system will be paralyzed, failing to achieve true autonomous collaboration and fault self-healing.

[0005] In recent years, breakthroughs in large language model technology have provided a new technical approach to solving the above problems. Its powerful natural language understanding, semantic reasoning, and knowledge transfer capabilities should have made it the best solution for addressing communication issues among heterogeneous devices in power distribution networks. However, the application of large language models in power systems is mainly concentrated in cloud-based non-real-time scenarios, with almost no application in edge real-time control. This is due to three main technical obstacles: First, centralized deployment suffers from excessive latency. The single-inference P99 latency of a 70B-level large language model is still over 87ms, while the protection response to emergency events such as voltage dips in the distribution network must be less than 20ms. The inference speed of cloud models is more than four times slower than the real-time control requirements. Second, there is a lack of dedicated semantic communication standards for the power industry. The interaction of large language models still uses natural language, which not only has low transmission efficiency but also easily leads to semantic ambiguity. Third, security risks cannot be effectively managed. Large language models have an inherent "illusion" problem, potentially outputting decisions that violate Kirchhoff's laws or equipment rated parameters in high-reliability scenarios such as power dispatching. Existing technologies rely solely on the model's own inference confidence for filtering, lacking a hard-constraint security verification mechanism independent of the model's inference path, making security risks uncontrollable.

[0006] In summary, existing technologies cannot simultaneously meet the requirements of power distribution networks for real-time performance, interoperability, security, and scalability. There is an urgent need for a solution that enables autonomous and collaborative control of heterogeneous devices at the edge, fundamentally addressing the inherent shortcomings of traditional centralized architectures and existing multi-agent solutions. Summary of the Invention

[0007] The purpose of this invention is to provide a multi-agent autonomous collaborative control method for the edge of a power distribution network, thereby solving all or one of the aforementioned problems in the prior art.

[0008] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0009] On the one hand, the present invention provides a multi-agent autonomous cooperative control method for the edge of a power distribution network, comprising the following steps:

[0010] Deploy lightweight large language models that can run locally on multiple edge device nodes of the power distribution network;

[0011] Construct a power-specific semantic primitive communication layer that is independent of the traditional power protocol layer;

[0012] When a triggering event occurs in the distribution network, the device agent corresponding to the triggering event acts as the source device agent, generates event information based on the triggering event, and broadcasts a semantic primitive message containing the event information to other device agents through the semantic primitive communication layer.

[0013] Other device agents that receive the semantic primitive message use a local lightweight large language model to perform semantic understanding and reasoning on the message, and generate a response message based on their own operating status;

[0014] The source device agent negotiates and generates candidate adjustment schemes based on the received response messages through a local lightweight large language model. The candidate adjustment schemes include at least one selected device agent and its corresponding adjustment task allocation.

[0015] The security verification module performs security verification on the candidate adjustment schemes. If the verification is successful, the candidate adjustment schemes are confirmed as executable adjustment schemes and sent for execution. Each selected device agent executes the assigned adjustment task and reports the execution result.

[0016] The distributed negotiation between edge device agents, the verification by the independent security verification module, and the task execution are all completed autonomously by the edge device agents without real-time control commands issued by the distribution network master station. When the source device agent confirms that the distributed negotiation cannot meet the adjustment requirements, or the security verification module continuously rejects the generated adjustment schemes to a preset number of times, it terminates the distributed negotiation and sends an alarm message to the distribution network master station to request the master station to perform centralized scheduling intervention.

[0017] As an improved approach, the default range of the broadcast is the communication domain composed of edge device agents within the same area; if the total adjustment capacity within the current communication domain cannot meet the adjustment requirements, the source device agent automatically expands the negotiation range to adjacent communication domains.

[0018] As an improved approach, after sensing a triggering event, the source device agent generates event information containing the total adjustment demand and response time requirements, and broadcasts it through the semantic primitive message; the response time requirements are determined by the source device agent based on the type of the triggering event and the local power grid safety operation rules.

[0019] As an improved approach, the process of semantic understanding and reasoning of the message using a local lightweight large language model is constrained by pre-embedded hard constraint rules for power grid security; these pre-embedded hard constraint rules for power grid security are embedded through at least one of the following mechanisms:

[0020] The power grid safety rules are encoded as non-trainable static tensors and fused with dynamic reasoning tensors at the attention mechanism layer.

[0021] During the model compression phase, knowledge distillation is performed on the grid topology connectivity and power balance equations to ensure that the lightweight model strictly follows Kirchhoff's laws when running on the edge side.

[0022] Establish a real-time feedback loop based on telemetry data to map the semantic decisions output by the model back to the physical simulation environment for pre-simulation. Only when the pre-simulation results are in the safe and stable domain are they allowed to be converted into control commands.

[0023] As an improved approach, the semantic primitive messages in the semantic primitive communication layer are encoded into a compact structured format, including a device unique identifier field, a message timestamp field, an intent type field, a parameter set field, a constraint set field, and a digital signature field; the structured format is either binary or Protobuf format; the total length of a single message does not exceed 256 bytes; the constraint set field includes upper and lower voltage limits, current protection settings, device rated power, and regulation response time requirements.

[0024] As an improved approach, the lightweight large language model is compressed to a parameter count of 100 million to 700 million through knowledge distillation, quantization pruning, and structured sparsity techniques, with a model size of no more than 1.5 GB. The lightweight large language model is deployed on ARM Cortex-A series or RISC-V series microcontrollers, and the parsing and inference latency of a single semantic primitive message is no more than 10 milliseconds.

[0025] As an improved approach, the response message is a moderating bid semantic primitive message containing intent, adjustable capacity, cost, and device reliability metrics; the source device agent performs negotiation based on all received moderating bid semantic primitive messages to generate a moderating scheme, specifically by performing multiple rounds of bidding negotiation, including:

[0026] The source device intelligence agent receives adjustment quotation messages from each device intelligence agent within a preset time window;

[0027] The semantic validity of each adjustment quotation message is evaluated. The semantic validity evaluation includes identifying whether the quotation content matches the intent type, whether the quotation parameters are complete, and whether the quotation value is within a reasonable range of the equipment's physical capabilities. Invalid quotations that fail the evaluation are ignored and recorded and are not included in subsequent combination optimization calculations.

[0028] After normalizing the bids of different dimensions and units among the valid bids that have passed the screening, the total adjustment demand, the response time requirement, the normalized adjustment cost of each valid bid, and the equipment reliability index are taken into account. A greedy algorithm is used to select equipment from high to low according to the comprehensive score until the adjustment demand is met, or a genetic algorithm is used for global combination optimization to generate the adjustment scheme.

[0029] If the total adjustment capacity of the current round is insufficient, the negotiation scope will be automatically expanded to the adjacent communication domain; if it still cannot be satisfied after the preset maximum number of rounds, the negotiation will be terminated and an alarm message will be sent to the distribution network master station.

[0030] As an improved approach, the triggering events include photovoltaic power fluctuations exceeding a preset threshold, load surges exceeding a preset threshold, and line faults; the edge device nodes include at least one of photovoltaic inverters, energy management systems for energy storage systems, charging pile controllers, flexible load controllers, static var generators, active power filters, distribution automation terminals, and microgrid central controllers.

[0031] As an improved solution, the security verification module performs security verification on the adjustment scheme independently of the inference process of the lightweight large language model, specifically including:

[0032] Based on a hard-coded power grid safety operation rule base, the adjustment scheme is matched and verified one by one. The rule base includes voltage upper and lower limit rules, current protection setting rules, equipment rated power limit rules, and power balance rules. Each rule is hard-coded and stored in a safe execution environment that is independent of the lightweight large language model inference path using an if-then structure.

[0033] If the verification fails, the adjustment scheme will be forcibly rejected regardless of the inference confidence of the lightweight large language model, and the source device agent will be triggered to re-execute the negotiation to generate a new adjustment scheme or security protection mechanism.

[0034] As an improved approach, the method further includes steps for traditional protocol compatibility and zero-sample protocol parsing:

[0035] Deploy protocol conversion agents on edge device nodes to achieve automatic bidirectional conversion between traditional power protocols and the semantic primitive communication layer through a lightweight large language model;

[0036] When an unknown binary message with an unrecognizable protocol header or data format is received, the protocol conversion agent performs zero-sample semantic reasoning based on the statistical characteristics of the message payload and the contextual temporal correlation, generates candidate semantic primitive messages and submits them for verification. After the verification is successful, the parsing mapping relationship is stored in the local protocol library.

[0037] The traditional power protocols include Modbus RTU, IEC 104, and DNP3 protocols; the automatic bidirectional conversion also includes converting semantic primitive messages into traditional power protocol binary messages and then sending them to existing devices that do not support semantic primitive communication.

[0038] The beneficial effects of the technical solution of this invention are:

[0039] The autonomous collaborative control method for multi-agent devices at the distribution network edge described in this invention can realize autonomous collaborative control of heterogeneous devices at the distribution network edge, reducing end-to-end response latency to milliseconds and effectively meeting the real-time control requirements for emergency events such as sudden drops in photovoltaic power and voltage dips. By constructing a power-specific semantic primitive communication layer, it fundamentally solves the "language incompatibility" problem caused by different protocols among heterogeneous devices, achieving semantic interoperability and plug-and-play functionality between devices. At the same time, through a dual protection mechanism of "secure embedding of the model inference layer + independent security verification of the decision layer," it effectively manages the decision illusion risk that may arise from large language models, ensuring that all control commands comply with the safe operation rules of the power grid. This method makes up for the inherent defects of traditional centralized control architectures and existing multi-agent solutions, and has high application value. Attached Figure Description

[0040] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0041] Figure 1 This is a system overall architecture diagram according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the semantic primitive message structure according to an embodiment of the present invention;

[0043] Figure 3 This is a flowchart of the distributed negotiation decision-making process according to an embodiment of the present invention;

[0044] Figure 4 This is a block diagram of the three-level security protection system according to an embodiment of the present invention. Detailed Implementation

[0045] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0046] In the description of this invention, it should be noted that the embodiments described in this invention are only some embodiments of this invention, not all embodiments; based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0047] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, apparatus, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0048] In the description of this invention, it should be noted that:

[0049] Lightweight large language models are large language models with 100 million to 700 million parameters. They are compressed through knowledge distillation, quantization pruning and structured sparsity techniques and can be deployed on ARM Cortex-A series, RISC-V series and microcontrollers with equivalent performance.

[0050] Semantic primitives are standardized semantic communication messages that are built independently of the traditional power protocol layer and contain an "intent-parameter-constraint" triple structure. They are used to achieve unified semantic interoperability between heterogeneous device agents.

[0051] A communication domain is a logical communication range consisting of multiple edge device agents within the same or adjacent stations, serving as the default boundary for agents to broadcast negotiation messages.

[0052] Hard constraint rules for power grid safety refer to the inviolable physical laws and safe operation rules such as Kirchhoff's current law, voltage upper and lower limits, and current protection settings, which are embedded into the reasoning process of a lightweight large language model by encoding them into untrainable static tensors or knowledge distillation.

[0053] The security verification module is a hard-coded rule verification unit that is independent of the lightweight large language model inference process. It is used to perform the final security compliance verification of all decision results.

[0054] The distribution network master station is the central system for distribution network operation monitoring and global scheduling. In this invention, it does not participate in the regular negotiation and decision-making process between edge device intelligent agents, but only intervenes in emergency scheduling as a fallback solution when distributed negotiation fails. It is also responsible for global status monitoring, data auditing and long-term optimization scheduling under normal conditions.

[0055] The multi-round bidding negotiation mechanism is a distributed negotiation process initiated by the source device intelligent agent, in which multiple device intelligent agents respond with quotations including adjustment power, response time and adjustment cost within a preset time window, and the source device intelligent agent selects the execution object according to the optimal combination.

[0056] Semantic understanding and reasoning refers to the process by which a lightweight large language model, after receiving a semantic primitive message, uses its natural language understanding and logical reasoning capabilities to extract and identify the task intent, key parameters, and constraints from the message. This process includes, but is not limited to, a standardized mapping and parsing method that maps different parameter representations from heterogeneous vendor devices to a local unified semantic space, and a direct generation method for deep intent recognition and response strategies that span multiple logical steps.

[0057] Example 1: This example provides a multi-agent autonomous collaborative control method for the edge of a power distribution network, designed to address the challenges of high response latency, poor interoperability of heterogeneous devices, and collaborative control issues arising from large-scale distributed energy access in traditional centralized control architectures. Figures 1-4 As shown, it includes:

[0058] S100, the steps for deploying edge agents and building the communication layer include:

[0059] S101. Deploy lightweight large language models that can run locally on multiple edge device nodes of the distribution network; the edge device nodes include at least one of photovoltaic inverters, energy management systems for energy storage systems, charging pile controllers, flexible load controllers, static var generators, active power filters, distribution automation terminals, and microgrid central controllers.

[0060] Preferably, this embodiment also provides a specific implementation scheme for a lightweight large language model suitable for edge devices:

[0061] Using the open-source Llama-2-7B-chat model from Meta as the base model, its parameter size was compressed from 7 billion to 300 million through a three-level compression technique:

[0062] (a) Knowledge distillation: The base model is distilled using a dataset containing 1 million power sector-specific data points, which covers power grid operation rules, equipment control instructions, fault handling cases and protocol parsing samples. The resulting student model has 300 million parameters.

[0063] (b) Quantization pruning: The student model is compressed using GPTQ 4-bit integer quantization technology, and 92% of redundant attention heads and feedforward layer parameters are removed through structured pruning;

[0064] (c) Structured sparsity: The feedforward layer and attention layer of the model are subjected to structured sparsity processing to further reduce computational complexity and memory usage.

[0065] After the three-level compression described above, the model size is approximately 1.2GB, which can be deployed on microcontrollers such as the ARM Cortex-A53@1.8GHz processor and the RISC-V RV64GC. Actual testing shows that the average parsing and inference time for a single semantic primitive message is 7.6ms, with a maximum of no more than 9.2ms, meeting the millisecond-level real-time control requirements at the edge of the power distribution network.

[0066] It should be noted that the compressed lightweight large language model mentioned above still retains the ability to understand and reason about unseen semantic combinations and control intentions within the target scenarios such as power regulation, fault response, and resource scheduling in power distribution networks. It does not degenerate into a rule matching system that only recognizes fixed patterns.

[0067] S102. Construct a power-specific semantic primitive communication layer independent of the traditional power protocol layer, and define a standardized set of semantic primitives containing a "intent-parameter-constraint" triple structure; wherein, the semantic primitive set covers the power regulation, fault response and resource scheduling business scenarios of the distribution network, and the semantic primitive messages are encoded in a structured format to ensure that the length of a single message does not exceed a preset number of bytes, such as 256 bytes.

[0068] S200, Event Triggering and Negotiation Initiation Steps, including:

[0069] S201. When a triggering event occurs in the distribution network, the intelligent agent of the source device corresponding to the triggering event generates event information based on the triggering event. The event information includes at least the total adjustment demand and response time requirements. The triggering event includes photovoltaic power fluctuations exceeding a preset threshold, load surges exceeding a preset threshold, and line faults.

[0070] S202, the source device agent broadcasts a semantic primitive message containing event information to all other device agents in its communication domain through the semantic primitive communication layer.

[0071] S300, capability resolution, and distributed negotiation steps include:

[0072] S301. Each device agent that receives the semantic primitive message performs semantic understanding and reasoning on the message using a local lightweight large language model, extracts the task intent and key parameters, and calculates its own adjustable capabilities by combining its real-time operating status and pre-embedded power grid safety hard constraint rules.

[0073] The semantic understanding and reasoning described herein is the core process by which a lightweight large language model performs deep parsing and intent understanding of received semantic primitive messages. In this embodiment, this process specifically includes, but is not limited to, the following implementation levels:

[0074] (a) Semantic normalization mapping parsing (preferred implementation):

[0075] After receiving semantic primitive messages, each device agent uses a local lightweight large language model to semantically parse the parameter set fields and constraint set fields in the messages. Because devices from different manufacturers express the same physical quantity differently in traditional protocols (for example, the adjustable power reported by an energy storage device may be called "available capacity," "remaining power," "maximum adjustable output," etc.), the lightweight large language model uses its semantic understanding capabilities to automatically identify and map these heterogeneous parameters into a locally preset unified semantic space, forming standardized key parameters such as "total adjustment demand" and "response time requirements" for subsequent calculations.

[0076] (b) Deep intent understanding and direct generation of response strategies (extended implementation):

[0077] For more complex negotiation scenarios, the receiving agent, when parsing messages, does not merely perform parameter mapping. Instead, it comprehensively understands the source device agent's intent, all constraints in the message, its own historical operational data, and its current real-time state. Leveraging the reasoning capabilities of a lightweight, large language model, it directly generates its optimal response strategy (including adjusting power levels, response time commitments, and cost assessments), skipping the explicit intermediate parameter mapping step. This approach offers a higher degree of "black box" control in semantic understanding and reasoning, but results in a faster overall response speed, making it suitable for scenarios where devices have established long-term collaborative relationships.

[0078] Regardless of the level used, the semantic understanding and reasoning process is constrained by pre-embedded hard constraints on power grid safety, ensuring that the reasoning results do not violate physical laws and safe operation boundaries. S302, each device agent replies to the source device agent via the semantic primitive communication layer with a regulation quotation message containing its own regulation power magnitude, regulation response time, and regulation cost;

[0079] S303. Based on all received regulation capability information, the source device agent performs distributed negotiation decision-making through a local lightweight large language model to generate candidate regulation schemes that meet regulation requirements, response time requirements, and optimal total cost, and assigns regulation tasks to selected device agents; specifically including:

[0080] S3031. The source device agent uses its local lightweight large language model to perform semantic validity verification on the adjustment quotation messages replied by each device agent. The semantic validity verification includes at least identifying whether the quotation content matches the intent type of the adjustment quotation message, whether the quotation parameters contain complete required fields, and whether the quotation value is within a reasonable range of the physical capabilities of the device agent, so as to filter out invalid quotations that are semantically incomplete or physically unreasonable.

[0081] After successful screening, the source device agent, using its local lightweight large language model, integrates the total adjustment demand, the response time requirement, the normalized adjustment cost of each valid bid, and equipment reliability indicators. It then invokes a preset optimization decision-making strategy to select equipment combinations from the valid bids that meet the total adjustment demand, generating the candidate adjustment scheme. The optimization decision-making strategy employs a greedy algorithm to select equipment sequentially from high to low comprehensive scores until the adjustment demand is met, or a genetic algorithm for global combination optimization. If the subsequent security verification stage rejects the candidate scheme and triggers renegotiation, the source device agent, based on the rejection reason and triggering rules fed back by the security verification module, incorporates the missing security constraints into the regeneration prompt and re-executes the negotiation decision. The maximum number of regenerations is preset to 3; for specific judgment and processing mechanisms, please refer to step S403.

[0082] The comprehensive score is obtained by the lightweight large language model according to the following multi-objective weighted calculation formula: Score_i = α·f(ResponseTime_i) +β·g(Cost_i)+γ·Reliability_i

[0083] Where ResponseTime_i is the adjustment response time (milliseconds) of the i-th device, Cost_i is the adjustment cost (yuan / kW) of the i-th device, Reliability_i is the adjustment success rate (0-1) of the i-th device in the past 30 days, α, β, and γ are preset weight coefficients, and f and g are normalization functions;

[0084] In the multi-round bidding negotiation mechanism, the start and end boundaries of each negotiation cycle are defined as follows: from the source device agent broadcasting a semantic primitive message containing event information via UDP until the expiration of a preset bid collection window; the bid collection window is of fixed duration, starting from when the source device agent sends the first broadcast message of this round, and the window duration is preset according to the type of power grid event (for example, for a voltage sag event, the window duration is set to 50ms). After the window expires, no new adjustment bids are received, and the process enters the scheme decision stage; if all known device agents in the communication domain complete their bid responses before the window expires, the window is not closed prematurely, but waits until the timer ends to ensure that all potential devices have the opportunity to participate in the bidding; the next round of negotiation starts with the source device agent sending a new broadcast message, and the timer restarts.

[0085] S3032. If the adjustment needs cannot be met in this round of negotiation, the source device agent automatically expands the negotiation scope, broadcasts the semantic primitive message containing event information to the device agents in adjacent communication domains, and receives a reply.

[0086] S3033. If the adjustment requirements still cannot be met after a preset maximum number of rounds of negotiation, the source device intelligent agent terminates the distributed negotiation process, generates and sends an alarm message to the distribution network master station to request the master station to perform centralized scheduling intervention. The preset maximum number of rounds is, for example, 3 rounds.

[0087] It should be noted that the centralized scheduling intervention is an emergency escalation mechanism triggered when distributed negotiation fails. Its priority and timeliness requirements are higher than the global optimization scheduling function executed by the main station on a daily basis. The two belong to different control levels.

[0088] S3034. If the negotiation is successful, the candidate adjustment scheme is output as the adjustment scheme to be verified for security.

[0089] S400, security verification and execution feedback steps, including:

[0090] S401. A security verification module independent of the lightweight large language model performs a final security verification on the adjustment scheme to be verified, specifically including:

[0091] S4011, Adjustment scheme for receiving the output of a lightweight large language model to be verified;

[0092] S4012. A hard-coded power grid safety operation rule base is used to perform matching and verification of adjustment schemes one by one. The rule base includes voltage upper and lower limit rules, current protection setting rules, equipment rated power limit rules and power balance rules. Each rule is hard-coded and stored in a safe execution environment that is independent of the lightweight large language model inference path using an if-then structure.

[0093] S4013. If the verification passes, the adjustment plan can be issued for execution; if the verification fails, the plan will be forcibly rejected regardless of how high the inference confidence of the large language model is.

[0094] S402. If the verification passes, the adjustment scheme to be verified is confirmed as the optimal executable adjustment scheme. Each selected device agent executes the assigned adjustment task and feeds back the execution result to the source device agent through the semantic primitive communication layer.

[0095] S403. If the verification fails, the adjustment scheme to be verified is rejected, and the process returns to step S3031. The source device agent then re-performs decision reasoning using a local lightweight large language model to generate a new candidate adjustment scheme. The maximum number of regenerations is preset to N_max (N_max ≥ 1, N_max = 3 in this embodiment), and the counter variable is k (initially k = 0). Each regeneration increments k by 1. If the scheme regenerated in the kth iteration (1 ≤ k ≤ N_max) passes the security verification, it is executed normally. If the scheme fails verification after N_max consecutive regenerations, a tiered security protection mechanism is triggered.

[0096] Level 1 (Immediate Response): Locks out the adjustment functions of relevant equipment, maintains the current operating state, and prevents system instability caused by erroneous commands;

[0097] Level 2 (Emergency Takeover): The source device intelligent agent immediately generates an emergency alarm message containing the reasons for continuous rejection, the current power grid status, and the record of the rejected schemes, and sends it to the distribution network master station, requesting the master station to initiate centralized dispatch intervention; the priority and timeliness requirements of this centralized dispatch intervention are higher than the global optimization dispatch function executed by the master station on a daily basis, which belongs to the emergency escalation processing mechanism, and is the same mechanism as the mechanism for requesting the master station to intervene when the distributed negotiation fails in step S3033.

[0098] For example, in a voltage over-limit regulation scenario, the source device's intelligent agent, after negotiation, generates a candidate regulation scheme: invoke the energy storage system to discharge at 500kW to support the voltage, while simultaneously instructing the photovoltaic inverter to reduce its active power output by 300kW. After receiving this scheme, the security verification module verifies it by matching each rule in a hard-coded rule base.

[0099] The first rule states: "The regulating power allocated to any device shall not exceed 1.0 times its rated power." The energy storage system has a rated power of 500kW, and the plan allocates 500kW, which has passed verification.

[0100] The second rule states: "The upper limit of the bus voltage is 1.05 pu." Based on the current grid topology and power injection changes, the digital twin simulation model shows that if the energy storage discharges at full power of 500kW, the bus voltage will rise to 1.08 pu, exceeding the safety limit, and the verification will fail.

[0101] The security verification module forcibly rejects the candidate solution and feeds back the reason for rejection and the trigger rule number to the source device agent. After receiving the feedback, the source device agent automatically incorporates "voltage upper limit 1.05pu" as a new hard constraint into the regeneration prompt and re-executes the negotiation decision. The new solution generated by the second negotiation is: the energy storage discharge power is adjusted to 350kW, and the static var generator is coordinated to absorb the remaining reactive power. After digital twin pre-simulation verification, the bus voltage is stabilized at 1.04pu. The solution is then executed after verification.

[0102] S500, result reporting and global monitoring steps include:

[0103] S501, The source device intelligent agent uploads the negotiation process data and execution results to the distribution network master station;

[0104] S502. The master station performs global operation status monitoring, data auditing and global optimization scheduling based on the uploaded data. The master station does not participate in the specific negotiation and decision-making process between the edge device intelligent agents in steps S200 to S400, nor does it issue specific control commands.

[0105] It should be clarified that the "master station not participating in decision-making" mentioned in this application specifically refers to a normal scenario where the distribution network is in normal operation or within the adjustable range of the edge. When edge distributed negotiation fails and triggers an alarm, the system automatically switches to "emergency escalation mode". In this mode, the master station acquires temporary control and issues mandatory scheduling instructions, which have higher priority than the local decisions of the edge agents. At this time, the edge agents unconditionally execute the master station's instructions until the system recovers to a safe and stable domain, after which it switches back to autonomous collaborative mode. Through the above-mentioned dual-mode switching mechanism of "normal autonomy and abnormal takeover", both low latency of edge control and system safety under extreme conditions are ensured.

[0106] It should be noted that, in this embodiment, the power-specific semantic primitive communication layer constructed in step S102 is specifically constructed as follows:

[0107] Extract the communication requirements of core business scenarios in the power distribution network and define basic intent types, including power reporting, power adjustment request, fault alarm, adjustment capability feedback, task confirmation, and execution result feedback;

[0108] Define a standardized set of parameters for each of the aforementioned basic intent types, the set of parameters including numerical parameters, time parameters, and state parameters;

[0109] Define power grid safety constraint parameters, including upper and lower voltage limits, current protection settings, equipment rated power, and regulation response time requirements;

[0110] Semantic primitive messages are encoded into a structured format that includes a device unique identifier field, a message timestamp field, an intent type field, a parameter set field, a constraint set field, and a digital signature field. The structured format includes, but is not limited to, JSON format, binary format, and Protobuf format.

[0111] It should be noted that in step S301 of this embodiment, the pre-embedded hard constraint rules for power grid security are embedded into the inference process of the lightweight large language model through the following mechanism:

[0112] The power grid safety rules are encoded as non-trainable static tensors and fused with dynamic inference tensors at the attention mechanism layer to ensure that key safety constraints are not covered by model parameter updates.

[0113] During the model compression phase, knowledge distillation is specifically performed on the grid topology connectivity and power balance equations to ensure that the lightweight model still strictly follows Kirchhoff's laws when running on the edge side.

[0114] Establish a real-time feedback loop based on telemetry data to map the semantic decisions output by the model back to the physical simulation environment for pre-simulation. Only when the pre-simulation results are in the safe and stable domain are they allowed to be converted into control commands.

[0115] Preferably, this embodiment also includes a traditional protocol compatibility and zero-shot protocol parsing step based on the lightweight large language model:

[0116] S601. Deploy a protocol conversion agent on the edge device node. The protocol conversion agent realizes automatic bidirectional conversion between traditional power protocols and the semantic primitive communication layer through a lightweight large language model.

[0117] S602, the protocol conversion agent receives binary messages of traditional power protocols (such as Modbus RTU messages, IEC 104 messages, DNP3 messages, etc.), parses the semantic meaning of the messages through a lightweight large language model—including identifying the meaning of function codes in the messages, the physical quantity type corresponding to the data address, and the dimensions and units of the data values—and generates corresponding semantic primitive messages before forwarding them to other agents.

[0118] S603. When an unknown binary message with an unrecognizable protocol header or data format is received, the lightweight large language model in the protocol conversion agent performs zero-sample semantic reasoning based on the statistical characteristics of the message payload and the contextual temporal correlation, automatically generating a corresponding candidate semantic primitive message, and submitting the candidate semantic primitive message to at least two other protocol conversion agents in the same communication domain for cross-validation; other protocol conversion agents compare and verify the operating status data of the same target device obtained at the same time through other known protocols (such as telemetry data of device electrical quantities). If the consistency between the physical quantity (such as voltage, current, power) in the candidate semantic primitive message and the telemetry data exceeds a preset threshold (such as 95%), the verification is successful, the parsing mapping relationship is stored in the local protocol library, and synchronized to the entire network; if the verification fails, the candidate semantic primitive message is discarded, and an "unknown protocol parsing failure" alarm is generated.

[0119] The zero-sample semantic reasoning refers to the lightweight large language model in the protocol conversion agent performing combined inference based on at least one of the following features, rather than relying on a pre-built protocol mapping table:

[0120] (a) Numerical distribution statistical characteristics: Analyze the numerical distribution pattern of each byte in the unknown message data payload, identify byte segments that conform to IEEE 754 floating-point, integer or BCD encoding mode, in order to locate the physical quantity field in the message;

[0121] (b) Timing correlation features: Perform timing analysis on multiple consecutive frames of messages from the same unknown message source to identify byte segments that change slowly (analog quantities, such as voltage and temperature) and jump (state quantities, such as switch positions) in order to infer the field type;

[0122] (c) Cross-protocol consistency verification feature: The candidate parsing results are compared with the electrical quantity data of the same target device obtained through a known protocol within the same time window, and the numerical consistency of the physical quantities of the power system under steady state is used for verification.

[0123] S604. The protocol conversion agent receives the semantic primitive message, generates the corresponding traditional power protocol binary message through a lightweight large language model, and then sends it to the existing devices that do not support semantic primitive communication, thereby achieving downlink compatibility with the existing devices.

[0124] Preferably, this embodiment further includes a semantic primitive expansion step:

[0125] S701. When new business scenarios or new equipment types appear in the distribution network (such as new energy storage equipment, V2G charging piles, hydrogen fuel cells, etc.), the lightweight large language model autonomously generates new semantic primitive suggestions based on its ability to understand new semantic combinations. The new semantic primitive suggestions include at least the intention type of the suggestions, the definition of the parameter set, the definition of the constraint set, and the corresponding business scenario description.

[0126] S702. The new semantic primitives are proposed to be uploaded to the distribution network master station, and standardized review is conducted by the master station operation and maintenance personnel or the automated review system.

[0127] S703. After the review is approved, the main station will uniformly synchronize the new semantic primitive definition to all intelligent agents covered by the semantic primitive communication layer, thereby realizing the online expansion of communication capabilities.

[0128] It should be noted that the semantic primitive expansion step is an offline / quasi-offline dictionary management process, which is different from the online real-time negotiation and decision-making process described in steps S200 to S400; the main station only participates in management at the semantic dictionary level and does not participate in real-time control decisions.

[0129] In summary, this solution enables intelligent agents to autonomously complete negotiation, communication, and control decisions at the edge, reducing end-to-end response latency to milliseconds and meeting the real-time control requirements of power distribution network emergencies. By constructing a semantic primitive communication layer, the "language barrier" problem caused by different protocols among heterogeneous devices is fundamentally solved, achieving semantic interoperability and plug-and-play functionality between devices. Simultaneously, through a dual protection mechanism of "secure embedding of the model inference layer + independent security verification of the decision layer," the risk of decision illusions that may arise from large language models is effectively managed, ensuring that all control commands comply with the rules for safe operation of the power grid.

[0130] Example 2: This example provides a multi-agent autonomous collaborative control system for the edge of a power distribution network, such as... Figure 1 As shown, it includes:

[0131] The deployment module is used to: deploy lightweight large language models that can run locally on multiple edge device nodes of the power distribution network; and build a power-specific semantic primitive communication layer that is independent of the traditional power protocol layer, defining a standardized set of semantic primitives containing the "intent-parameter-constraint" triple structure;

[0132] The negotiation initiation module is used to: determine the source device intelligent agent corresponding to the triggering event when a triggering event occurs in the distribution network; and control the source device intelligent agent to broadcast a semantic primitive message containing event information to all other device intelligent agents in its communication domain through the semantic primitive communication layer.

[0133] The negotiation and decision-making module is used to: control each device agent receiving the semantic primitive message to parse the intent and parameters of the message using a local lightweight large language model, and calculate its own available regulation capacity and regulation cost by combining its own real-time operating status and pre-embedded hard constraint rules for power grid safety; control each device agent to reply to the source device agent with a regulation quotation message containing the regulation power magnitude, regulation response time, and regulation cost; and control the source device agent to perform semantic-level understanding and comprehensive decision-making reasoning on the regulation quotation messages replied by each device agent through its local lightweight large language model, and make a decision based on the total regulation demand, response time, and regulation cost. Based on time requirements and the principle of optimal total cost, candidate adjustment schemes are autonomously generated, and adjustment tasks are assigned to selected device agents. The decision-making and reasoning process of the lightweight large language model can be embedded with logical constraints of greedy algorithms or genetic algorithms to guide the model to reason and generate schemes within a semantic space that meets the requirements of optimal solution. If the adjustment requirements cannot be met in this round of negotiation, the source device agent is controlled to automatically broadcast the semantic primitive message containing event information to adjacent communication domains to expand the negotiation scope. If the requirements still cannot be met after multiple rounds of negotiation, an alarm message is generated and reported to the master station to request centralized scheduling intervention.

[0134] The security verification and execution module is used to: perform a final security verification on the candidate adjustment schemes by a security verification module independent of the lightweight large language model, to verify whether they comply with the power grid safe operation rules and filter out decisions that do not comply with the rules; if the verification passes, the candidate adjustment scheme is confirmed as the optimal executable adjustment scheme, and each selected device agent is controlled to execute the assigned adjustment task, and the execution result is fed back to the source device agent through the semantic primitive communication layer; if the verification fails, the candidate adjustment scheme is rejected, and the negotiation decision module is triggered, so that the source device agent re-performs decision reasoning through the local lightweight large language model to generate a new candidate adjustment scheme; if the verification still fails after multiple regenerations, a security protection mechanism is triggered.

[0135] The reporting and monitoring module is used to control the source device intelligent agent to upload the negotiation process data and execution results to the distribution network master station, so that the master station can perform global operation status monitoring, data auditing and global optimization scheduling based on the uploaded data. The master station does not participate in the specific negotiation and decision-making process between the edge device intelligent agents in the negotiation initiation module, negotiation decision module and security verification and execution module.

[0136] Example 3: This example provides a computer-readable storage medium, including:

[0137] The storage medium is used to store computer software instructions used to implement the autonomous collaborative control method of the multi-agent distribution network edge described in Embodiment 1 above. It includes programs for executing the above-mentioned programs set for the autonomous collaborative control method of the multi-agent distribution network edge. Specifically, the executable program can be built into the autonomous collaborative control system of the multi-agent distribution network edge described in Embodiment 2. In this way, the autonomous collaborative control system of the multi-agent distribution network edge can implement the autonomous collaborative control method of the multi-agent distribution network edge described in Embodiment 1 by executing the built-in executable program.

[0138] Furthermore, the computer-readable storage medium in this embodiment can be any combination of one or more readable storage media, wherein the readable storage medium includes an electrical, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0139] Unlike existing technologies, this application presents a multi-agent autonomous collaborative control method for distribution network edges, enabling millisecond-level autonomous collaborative control of heterogeneous devices at the distribution network edge. This solves the challenges of high response latency, poor interoperability of heterogeneous devices, and large-scale distributed energy access issues inherent in traditional centralized architectures. By deploying a lightweight large language model on edge devices and constructing a power-specific semantic primitive communication layer, cross-protocol and cross-vendor device semantic interoperability and plug-and-play functionality are achieved. Simultaneously, a dual protection mechanism—secure embedding at the model inference layer and independent security verification at the decision layer—effectively manages the decision illusion risk of the large language model in power control scenarios, ensuring that all control commands strictly comply with power grid safety operation rules. This significantly improves the distribution network's capacity to absorb distributed energy and its fault self-healing capabilities, overcoming the shortcomings of existing technologies and demonstrating high application value.

[0140] It should be understood that in the various embodiments of this document, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this document.

[0141] It should also be understood that, in the embodiments herein, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following associated objects have an "or" relationship.

[0142] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this document.

[0143] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0144] In the embodiments provided herein, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, devices, or units, or they may be electrical, mechanical, or other forms of connection.

[0145] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments described herein, depending on actual needs.

[0146] Furthermore, the functional units in the various embodiments of this document can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this paper, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this paper. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A power distribution network edge multi-agent autonomous collaborative control method, characterized in that, Includes the following steps: Deploy lightweight large language models that can run locally on multiple edge device nodes of the power distribution network; Construct a semantic primitive communication layer that is independent of the traditional power protocol layer and only carries the interaction of power business semantic primitives; When a triggering event occurs in the distribution network, the device agent corresponding to the triggering event acts as the source device agent, generates event information based on the triggering event, and broadcasts a semantic primitive message containing the event information to other device agents through the semantic primitive communication layer. Other device agents that receive the semantic primitive message use a local lightweight large language model to perform semantic understanding and reasoning on the message, and generate a response message based on their own operating status; The source device agent negotiates and generates candidate adjustment schemes based on the received response messages through a local lightweight large language model. The candidate adjustment schemes include at least one selected device agent and its corresponding adjustment task allocation. The security verification module performs security verification on the candidate adjustment schemes. If the verification is successful, the candidate adjustment schemes are confirmed as executable adjustment schemes and sent for execution. Each selected device agent executes the assigned adjustment task and reports the execution result. The distributed negotiation between edge device agents, the verification by the security verification module, and the task execution are all completed autonomously by the edge device agents without real-time control commands issued by the distribution network master station. When the source device agent confirms that the distributed negotiation cannot meet the adjustment requirements, or the security verification module continuously rejects the generated adjustment schemes to a preset number of times, it terminates the distributed negotiation and sends an alarm message to the distribution network master station to request the master station to perform centralized scheduling intervention.

2. The power distribution grid edge multi-agent autonomous cooperative control method according to claim 1, characterized in that: The default range of the broadcast is the communication domain composed of device agents within the same area; if the total adjustment capacity within the current communication domain cannot meet the adjustment requirements, the source device agent will automatically expand the negotiation range to adjacent communication domains.

3. The multi-agent autonomous cooperative control method for the distribution network edge according to claim 1, characterized in that: After sensing a triggering event, the source device agent generates event information containing the total adjustment demand and response time requirements, and broadcasts it through the semantic primitive message; the response time requirements are determined by the source device agent based on the type of triggering event and the local power grid safety operation rules.

4. The multi-agent autonomous cooperative control method for the distribution network edge according to claim 1, characterized in that: The process of semantic understanding and reasoning of the message using a local lightweight large language model is constrained by pre-embedded hard constraint rules for power grid security; the pre-embedded hard constraint rules for power grid security are embedded through at least one of the following mechanisms: The power grid safety rules are encoded as non-trainable static tensors and fused with dynamic reasoning tensors at the attention mechanism layer. During the model compression phase, knowledge distillation is performed on the grid topology connectivity and power balance equations to ensure that the lightweight model strictly follows Kirchhoff's laws when running on the edge side. Establish a real-time feedback loop based on telemetry data to map the semantic decisions output by the model back to the physical simulation environment for pre-simulation. Only when the pre-simulation results are in the safe and stable domain are they allowed to be converted into control commands.

5. The multi-agent autonomous cooperative control method for the distribution network edge according to claim 1, characterized in that: The semantic primitive messages in the semantic primitive communication layer are encoded in a compact structured format, including a device unique identifier field, a message timestamp field, an intent type field, a parameter set field, a constraint set field, and a digital signature field; the structured format is a binary format or a Protobuf format; the total length of a single message does not exceed 256 bytes; the constraint set field includes upper and lower voltage limits, current protection settings, device rated power, and regulation response time requirements.

6. The multi-agent autonomous cooperative control method for the distribution network edge according to claim 1, characterized in that: The lightweight large language model is compressed to a parameter count of 100 million to 700 million through knowledge distillation, quantization pruning, and structured sparsity techniques, with a model size of no more than 1.5 GB. The lightweight large language model is deployed on ARM Cortex-A series or RISC-V series microcontrollers, and the parsing and inference latency of a single semantic primitive message is no more than 10 milliseconds.

7. The multi-agent autonomous cooperative control method for the distribution network edge according to claim 3, characterized in that: The response message is an adjustment quotation semantic primitive message containing intent, adjustable capacity, cost, and equipment reliability indicators; The source device agent performs negotiation based on all received adjustment bid semantic primitive messages to generate an adjustment scheme, specifically by performing multiple rounds of bidding negotiation, including: The source device intelligence agent receives adjustment quotation messages from each device intelligence agent within a preset time window; The semantic validity of each adjustment quotation message is evaluated. The semantic validity evaluation includes identifying whether the quotation content matches the intent type, whether the quotation parameters are complete, and whether the quotation value is within a reasonable range of the equipment's physical capabilities. Invalid quotations that fail the evaluation are ignored and recorded and are not included in subsequent combination optimization calculations. After normalizing the bids of different dimensions and units among the valid bids that have passed the screening, the total adjustment demand, the response time requirement, the normalized adjustment cost of each valid bid, and the equipment reliability index are taken into account. A greedy algorithm is used to select equipment from high to low according to the comprehensive score until the adjustment demand is met, or a genetic algorithm is used for global combination optimization to generate the adjustment scheme. If the total adjustment capacity of the current round is insufficient, the negotiation scope will be automatically expanded to the adjacent communication domain; if it still cannot be satisfied after the preset maximum number of rounds, the negotiation will be terminated and an alarm message will be sent to the distribution network master station.

8. The multi-agent autonomous cooperative control method for the distribution network edge according to claim 1, characterized in that: The triggering events include photovoltaic power fluctuations exceeding a preset threshold, load surges exceeding a preset threshold, and line faults; the edge device nodes include at least one of photovoltaic inverters, energy storage system energy management systems, charging pile controllers, flexible load controllers, static var generators, active power filters, distribution automation terminals, and microgrid central controllers.

9. The multi-agent autonomous cooperative control method for the distribution network edge according to claim 1, characterized in that: The security verification module, independent of the inference process of the lightweight large language model, performs security verification on the adjustment scheme, specifically including: Based on a hard-coded power grid safety operation rule base, the adjustment scheme is matched and verified one by one. The rule base includes voltage upper and lower limit rules, current protection setting rules, equipment rated power limit rules, and power balance rules. Each rule is hard-coded and stored in a safe execution environment that is independent of the lightweight large language model inference path using an if-then structure. If the verification fails, the adjustment scheme will be forcibly rejected regardless of the inference confidence of the lightweight large language model, and the source device agent will be triggered to re-execute the negotiation to generate a new adjustment scheme or security protection mechanism.

10. The multi-agent autonomous cooperative control method for the edge of a power distribution network according to claim 1, characterized in that: The method also includes steps for traditional protocol compatibility and zero-sample protocol parsing: Deploy protocol conversion agents on edge device nodes to achieve automatic bidirectional conversion between traditional power protocols and the semantic primitive communication layer through a lightweight large language model; When an unknown binary message with an unrecognizable protocol header or data format is received, the protocol conversion agent performs zero-sample semantic reasoning based on the statistical characteristics of the message payload and the contextual temporal correlation, generates candidate semantic primitive messages and submits them for verification. After the verification is successful, the parsing mapping relationship is stored in the local protocol library. The traditional power protocols include Modbus RTU, IEC 104, and DNP3 protocols; the automatic bidirectional conversion also includes converting semantic primitive messages into traditional power protocol binary messages and then sending them to existing devices that do not support semantic primitive communication.

Images